CSE 544Principles of DatabaseManagement Systems

Magdalena BalazinskaFall 2007

Lecture 14 - Replication

CSE 544 - Fall 2007 2

References

• The Dangers of Replication and a Solution.Jim Gray, Pat Helland, Patrick O’Neil, and DennisShasha. SIGMOD’96.

• Database management systems.Ramakrishnan and Gehrke.Third Ed. Chapter 22.11

CSE 544 - Fall 2007 3

Outline

• Goals of replication

• Three types of replication– Eager replication– Lazy replication– Two-tier replication

• Replication support in commercial databases

CSE 544 - Fall 2007 4

Goals of Replication

• Goal 1: availability• Goal 2: performance

• But, it’s easy to build a replicated system that reducesperformance and availability

Somerequests

Otherrequests

Three replicas

CSE 544 - Fall 2007 5

Eager Replication

• Also called synchronous replication• All updates are applied to all replicas (or to a majority) as

part of a single transaction (need two phase commit)• Main goal: as if there was only one copy

– Maintain consistency– Maintain one-copy serializability– I.e., execution of transactions has same effect as an execution on

a non-replicated db

• Transactions must acquire global locks

CSE 544 - Fall 2007 6

Eager Master

• One master for each object holds primary copy– To update object, transaction must acquire a lock at the master– Lock at the master is global lock

CSE 544 - Fall 2007 7

Crash Failures

• What happens when a secondary crashes?– Nothing happens– When secondary recovers, it catches up

• What happens when the master/primary fails?– Blocking would hurt availability– Must chose a new primary: run election

CSE 544 - Fall 2007 8

Network Failures

• Network failures can cause trouble...– Secondaries think that primary failed– Secondaries elect a new primary– But primary can still be running– Now have two primaries!

CSE 544 - Fall 2007 9

Majority Consensus

• To avoid problem, only majority partition can continueprocessing at any time

• In general,– Whenever a replica fails or recovers...– a set of communicating replicas must determine...– whether they have a majority before they can continue

CSE 544 - Fall 2007 10

Eager Group

• With n copies– Exclusive lock on x copies is global exclusive lock– Shared lock on s copies is global shared lock– Must have: 2x > n and s + x > n

• Majority locking– s = x = (n+1)/2– No need to run any reconfiguration algorithms

• Read-locks-one, write-locks-all– s=1 and x = n, high read performance– Need to make sure this algorithm runs on a quorum of computers

CSE 544 - Fall 2007 11

Eager Replication Properties

• Favours consistency over availability– Only majority partition can process requests– There appears to be a single copy of the db

• High runtime overhead– Must lock and update at least majority of replicas– Two-phase commit– Runs at pace of slowest replica in quorum– So overall system is now slower– Higher deadlock rate (transactions take longer)

CSE 544 - Fall 2007 12

Lazy Replication

• Also called asynchronous replication• Also called optimistic replication

• Main goals: availability and performance

• Approach– One replica updated by original transaction– Updates propagate asynchronously to other replicas

CSE 544 - Fall 2007 13

Lazy Master

• One master holds primary copy– Transactions update primary copy– Master asynchronously propagates updates to replicas, which

process them in same order– Ensures single-copy serializability

• What happens when master/primary fails?– Can lose most recent transactions when primary fails!– After electing a new primary, secondaries must agree who is

most up-to-date

CSE 544 - Fall 2007 14

Lazy Group

• Also called multi-master• Best scheme for availability• Cannot guarantee one-copy serializability!

R1 R2Init: x=1Update x=2

Init: x=1Update x=3

CSE 544 - Fall 2007 15

Lazy Group

• Cannot guarantee one-copy serializability!• Instead guarantee convergence

– Db state does not reflect any serial execution– But all replicas have the same state

• Detect conflicts and reconcile replica states• Different reconciliation techniques are possible

– Manual– Most recent timestamp wins– Site A wins over site B– User-defined rules, etc.

CSE 544 - Fall 2007 16

Detecting ConflictsUsing TimestampsR1 R2

x=2, Old: T0 New: T1

Init: x=1 at T0Update at T1 : x=2

Init: x=1 at T0

x=2 at T1x=2 at T1

CSE 544 - Fall 2007 17

R1 R2

x=2, Old: T0 New: T1

Conflict!

Init: x=1 at T0Update at T1 : x=2

Init: x=1 at T0

Update at T2: x=3

Conflict! x=3, Old: T0 New: T2

Reconciliation ruleT2 > T1, so x=3

Reconciliation ruleT2 > T1, so x=3

Detecting ConflictsUsing Timestamps

CSE 544 - Fall 2007 18

Lazy GroupReplication Properties

• Favours availability over consistency– Can read and update any replica– High runtime performance

• Weak consistency– Conflicts and reconciliation

CSE 544 - Fall 2007 19

Two-Tier Replication

• Benefits of lazy master and lazy group• Each object has a master with primary copy• When disconnected from master

– Secondary can only run tentative transactions

• When reconnects to master– Master reprocesses all tentative transactions– Checks an acceptance criterion– If passes, we now have final commit order– Secondary undoes tentative and redoes committed

CSE 544 - Fall 2007 20

Replication inCommercial DBMSs

• All major vendors provide replication– Oracle 9i, DB2, SQL Server

• Eager master• Lazy master• Lazy group

– Timestamps or before-values to detect conflicts– Pre-defined rules for state reconciliation– Support for user-defined reconciliation rules

CSE 544 - Fall 2007 21

Conclusion• Replication is a very important problem

– Fault-tolerance (various forms of replication)– Caching (lazy master)– Warehousing (lazy master)– Mobility (two-tier techniques)

• Replication is complex, but basic techniques andtrade-offs are very well known

– Eager or lazy replication– Master or no master– For eager replication: use quorum